Add READMEs to featured environments

environments/featured/browser-use/Dockerfile.hud

@@ -2,7 +2,7 @@ FROM mcr.microsoft.com/playwright/python:v1.63.0-noble@sha256:72bd171a9ffc2b4b59
 WORKDIR /app
 COPY requirements.txt /app/
 COPY .hud-sdk/ .hud-sdk/
-ARG HUD_SDK_VERSION=0.6.19
+ARG HUD_SDK_VERSION=0.6.20
 RUN python3 -m pip install --no-cache-dir --require-hashes -r requirements.txt \
     && python3 -c "from importlib.metadata import version; import os; assert version('hud') == os.environ['HUD_SDK_VERSION']; assert version('playwright') == '1.63.0'; print('HUD_ENV_SDK_VERSION=' + version('hud'))"
 COPY --chmod=644 env.py tasks.py /app/

environments/featured/browser-use/README.md

@@ -0,0 +1,20 @@
+# Browser Use
+
+A visual web-research environment backed by a real Chromium browser. Agents can
+inspect screenshots, navigate, click, scroll, and use the browser's developer
+protocol without leaving the allowed website.
+
+The featured task asks the agent to inspect rental photos and count the chairs
+at a specific kitchen table. Grading checks both the exact answer and whether
+the agent cited screenshot evidence that actually showed the kitchen.
+
+The environment exposes:
+
+- `cdp`: direct browser control for compatible agents
+- `mcp`: screenshot and browser-interaction tools
+
+Deploy from this directory with:
+
+```bash
+hud deploy
+```

environments/featured/browser-use/requirements.in

@@ -1,3 +1,3 @@
-hud[agents]==0.6.19
+hud[agents]==0.6.20
 playwright==1.63.0
 uvicorn==0.54.0

environments/featured/browser-use/requirements.txt

@@ -1,5 +1,5 @@
 # This file was autogenerated by uv via the following command:
-#    uv --no-config pip compile requirements.in --generate-hashes --universal --python-version 3.12 -o requirements.txt --offline
+#    uv --no-config pip compile requirements.in --generate-hashes --universal --python-version 3.12 -o requirements.txt
 aiofile==3.12.3 \
     --hash=sha256:5c1bcc9e929c50834608e8cc1a4cc1d7503eb60c15a535b779fd39e2f372c017 \
     --hash=sha256:caa6aa746b5e47e2165f7abd741b6415e49cf4d44fddc0f61844612cc3924d41
@@ -618,9 +618,9 @@ httpx2-jsfetch==1.0 ; sys_platform == 'emscripten' \
     --hash=sha256:70a0e3eabfef7cce5ad9c629f7d01ca05e418f586646f4ddf14782e4c1454c60 \
     --hash=sha256:cb916b707601e69a07721aabc8f3f6659be3a6893bc1ff5c6f9e02241df2da32
     # via httpx2
-hud==0.6.19 \
-    --hash=sha256:3288dc887ac4489383bb15ddcc10d3f59440b4af9fbfbed22c44598dffa00c94 \
-    --hash=sha256:f9c45d614bb2efac228d41f23f215a1ae874bbe87eb73777ac90d994af2519f7
+hud==0.6.20 \
+    --hash=sha256:4ffb877da4e9fd7f0c397fb60874a997824c110d3e6903fae2595f35b81a66f9 \
+    --hash=sha256:7f5d084c12093dfe68b4e845cd428cf29208545d57085b6bb7c8c11437ec6802
     # via -r requirements.in
 idna==3.20 \
     --hash=sha256:a7db850025b95ded1eae8a46181a1a6c56c92c96f0e2b005d9ff8dc0210cab44 \

environments/featured/file-analysis/Dockerfile.hud

@@ -5,7 +5,7 @@ WORKDIR /hud
 COPY pyproject.toml uv.lock env.py ./
 COPY fixtures ./fixtures
 COPY .hud-sdk/ .hud-sdk/
-ARG HUD_SDK_VERSION=0.6.19
+ARG HUD_SDK_VERSION=0.6.20
 RUN UV_PROJECT_ENVIRONMENT=/usr/local/venv uv sync --frozen --no-install-project --no-dev && chmod -R a+rX /usr/local/venv && chmod -R go-rwx /hud
 ENV PATH="/usr/local/venv/bin:${PATH}" WORKSPACE_DIR=/workspace
 RUN python -c "from importlib.metadata import version; import os; assert version('hud') == os.environ['HUD_SDK_VERSION']; print('HUD SDK', version('hud'))"

environments/featured/file-analysis/README.md

@@ -0,0 +1,21 @@
+# File Analysis
+
+A document-analysis environment for tasks built from prompts, uploaded files,
+and weighted grading criteria. Files are staged in an isolated workspace where
+the agent can inspect them with shell tools and write a final report.
+
+The featured task is a supplier-payment reconciliation. It combines invoices,
+bank payments, receiving records, and vendor terms, then grades the resulting
+report for correct calculations, evidence, and recommended follow-up.
+
+The environment provides:
+
+- an isolated shell workspace with network access disabled
+- file tracking for the agent's report
+- criteria-based grading with an LLM judge
+
+Deploy from this directory with:
+
+```bash
+hud deploy
+```

environments/featured/file-analysis/pyproject.toml

@@ -4,7 +4,7 @@ version = "0.1.0"
 description = "Argument-hints environment: prompt, data-file attachments, and rubric grading via console form hints (HUD v6)"
 requires-python = ">=3.12,<3.13"
 dependencies = [
-    "hud[agents]==0.6.19",
+    "hud[agents]==0.6.20",
     "httpx>=0.27.0",
 ]
 

environments/featured/flask-coding/Dockerfile.hud

@@ -9,7 +9,7 @@ WORKDIR /hud
 COPY pyproject.toml uv.lock env.py grader.py flask.bundle ./
 
 COPY .hud-sdk/ .hud-sdk/
-ARG HUD_SDK_VERSION=0.6.19
+ARG HUD_SDK_VERSION=0.6.20
 RUN UV_PROJECT_ENVIRONMENT=/usr/local/venv uv sync --frozen --no-install-project \
   && /usr/local/venv/bin/python -c "from importlib.metadata import version; import os; assert version('hud') == os.environ['HUD_SDK_VERSION']; print('HUD_SDK_VERSION=' + version('hud'))" \
   && mkdir /app \

environments/featured/flask-coding/README.md

@@ -0,0 +1,20 @@
+# Flask Coding
+
+A coding environment that gives the agent an isolated Flask repository and
+grades its changes with hidden tests. Each task starts from a clean git
+baseline, exposes only the working tree to the agent, and applies the hidden
+test patch after the agent finishes.
+
+The featured task asks the agent to add binary-file loading support to
+`flask.Config.from_file` while preserving the existing text and error-handling
+behavior. The score comes from targeted fail-to-pass and regression tests.
+
+The workspace blocks network access and runs the agent as an unprivileged user.
+Existing runs remain reproducible because each task pins its repository
+baseline and test contract.
+
+Deploy from this directory with:
+
+```bash
+hud deploy
+```

environments/featured/flask-coding/pyproject.toml

@@ -6,7 +6,7 @@ requires-python = ">=3.11,<3.13"
 dependencies = [
     "blinker==1.5",
     "click>=8.0",
-    "hud==0.6.19",
+    "hud==0.6.20",
     "itsdangerous>=2.0",
     "jinja2>=3.0",
     "pytest>=8.0,<9",

environments/featured/libbet/Dockerfile.hud

@@ -14,7 +14,7 @@ WORKDIR /app
 # the lock, and the runtime modules the wheel build needs (pyproject `only-include`).
 COPY pyproject.toml uv.lock env.py tasks.py emulator.py scoring.py games_loader.py ./
 COPY .hud-sdk/ .hud-sdk/
-ARG HUD_SDK_VERSION=0.6.19
+ARG HUD_SDK_VERSION=0.6.20
 RUN uv sync --frozen --no-dev
 RUN /app/.venv/bin/python -c "import importlib.metadata; import os; version = importlib.metadata.version('hud'); print('HUD_SDK_VERSION=' + version); assert version == os.environ['HUD_SDK_VERSION']"
 

environments/featured/libbet/README.md

@@ -0,0 +1,19 @@
+# Libbet
+
+A visual Game Boy environment built from the freely licensed homebrew puzzle
+game *Libbet and the Magic Floor*. Agents see the current screen and play with
+MCP tools for the directional pad, buttons, waiting, and screenshots.
+
+The emulator advances only when a tool is called, so the game stays paused
+while the agent reasons. The featured task gives the agent 40 action steps to
+explore the game.
+
+Scoring rewards visually distinct progress discovered during play rather than
+requiring a particular level completion. Screenshots and scoring details are
+recorded in the trace.
+
+Deploy from this directory with:
+
+```bash
+hud deploy
+```

environments/featured/libbet/pyproject.toml

@@ -4,7 +4,7 @@ version = "0.1.0"
 description = "A HUD environment template that wraps VideoGameBench Game Boy games as scored, turn-based agent evals."
 requires-python = ">=3.11,<3.13"
 dependencies = [
-    "hud==0.6.19",
+    "hud==0.6.20",
     "pyboy>=2.7.0",          # Game Boy backend
     "imagehash>=4.3.1",
     "pillow>=10.0.0",

environments/featured/robot-gripper/.dockerignore

@@ -22,3 +22,4 @@ examples/
 docs/
 run_vla.py
 *.md
+!README.md

environments/featured/robot-gripper/Dockerfile.hud

@@ -28,7 +28,7 @@ WORKDIR /app
 COPY pyproject.toml uv.lock ./
 COPY wheels/ wheels/
 COPY .hud-sdk/ .hud-sdk/
-ARG HUD_SDK_VERSION=0.6.19
+ARG HUD_SDK_VERSION=0.6.20
 RUN uv sync --frozen --no-dev
 RUN /app/.venv/bin/python -c "from importlib.metadata import version; import os; assert version('hud') == os.environ['HUD_SDK_VERSION']; print('HUD SDK:', version('hud'))"
 

environments/featured/robot-gripper/README.md

@@ -0,0 +1,19 @@
+# Robot Gripper
+
+A tabletop robotics environment powered by a Newton physics simulation. Agents
+control a gripper through MCP tools, inspect rendered views and simulator state,
+and receive rewards from the resulting physical state.
+
+The environment includes tasks for opening a drawer, lifting and moving an
+object, and maintaining a force-controlled grasp. The featured task asks the
+agent to pick up a mug within a 40-step budget.
+
+Grading measures continuous progress as well as task completion, so partial
+movement receives a proportional score while reaching the target earns the
+success bonus.
+
+Deploy from this directory with:
+
+```bash
+hud deploy
+```

environments/featured/robot-gripper/pyproject.toml

@@ -7,7 +7,7 @@ requires-python = ">=3.12,<3.13"
 # Eval-machine dependencies. The Newton physics engine (`newton` + `gizmo`) is
 # NOT on PyPI: it ships as a pre-built wheel in wheels/ (see [tool.uv.sources]).
 dependencies = [
-    "hud==0.6.19",  # HUD SDK: Environment, capabilities, hud eval
+    "hud==0.6.20",  # HUD SDK: Environment, capabilities, hud eval
     "mujoco>=3.2.0",       # rendering, model name lookups, IK
     "numpy>=1.26.0",
     "Pillow>=10.0.0",      # base64 PNG encode/decode for render()